Chapter 5
Dynamic Sliding ModeControl and OutputFeedback
C. EDWARDS and S.K. SPURGEONUniversity of Leicester, England, United Kingdom
5.1 Introduction
The sliding mode design approach involves two distinct stages. The firstconsiders the design of a switching function which provides desirable systemperformance in the sliding mode. The second consists of designing a controllaw which will ensure the sliding mode, and thus the desired performance,is attained and maintained. The first stage is often termed the existenceproblem and the second the reachability problem. Traditionally much of thework in the area of sliding mode control considered uncertain, often linear,state-space systems and the solution of both the existence and reachabilityproblems assumed full state information was available to the control law.Thus, a switching function would be determined that was a function of thesystem states and an associated state -dependent control law would result.
Clearly the assumption of full state availability is restrictive; it may beimpossible or impractical to measure all the states for many processes. Onepossible solution is to use an observer to estimate the system states andsliding mode techniques for such observation have been illustrated in theprevious chapter. The alternative is to consider solutions to the existenceand reachability problems which are dependent on system outputs alone.
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
Uncertain linear systems, represented by a nominal (A, B, C] triple, will bethe initial focus of this chapter. A straightforward solution to the problemof sliding mode control via output feedback will be seen to be possibleif the nominal triple is relative degree one, i.e., the product CB is fullrank, and the transmission zeros of the nominal system are in the lefthand plane, i.e., the triple is minimum phase. As may be expected fromthe full state scenario, these transmission zeros will appear as poles of thedynamics in the sliding mode. In fact, when the number of outputs andinputs is equal, these transmission zeros will wholly determine the slidingmode performance in general and the existence problem is trivial. If thereare more measured process outputs than control inputs, then it will beseen that the solution to the existence problem may be formulated as thedesign of a static output feedback controller for a particular sub-systemtriple. It is well known that any triple is stabilizable via static outputfeedback if it is both controllable and observable and satisfies a certaininequality which is a function of the system dimensions. This latter resultis often termed the Kimura-Davison Condition. It will be shown thata sufficient condition to solve the existence problem can be formulated.This depends on the satisfaction of a similar inequality relating to thesystem dimensions and the number of transmission zeros of the originaltriple. If this inequality does not hold for the process of interest, then theexistence problem can always be solved by introducing a compensator. Thiseffectively amounts to augmenting the system with some extra dynamicsthat are driven by the outputs of the plant. In this case it will be seenthat the existence problem and the design of the compensator may beeffectively accomplished by solving a particular output feedback problem.Here the inequality which must be satisfied will be seen to relate to thedimension of the compensator as well as the dimensionality and number oftransmission zeros of the system. The first is a design variable which willensure that a switching function can be found to make the sliding motionstable. This chapter will go on to present output dependent reachabilityconditions that will ensure that the sliding mode is ultimately attractiveand that the designed dynamics are attained.
It has been seen in the above that the use of dynamic feedback is desir-able to broaden the class of linear systems for which sliding mode controllersdependent only on system outputs may be designed. A second area wherethe use of dynamic feedback yields useful properties is in the sliding modecontrol of nonlinear systems. The results described above are only appli-cable where the process of interest may be modelled by a linear uncertainsystem. However, some processes are so nonlinear that such a modellingassumption is invalid. Many results in the literature in the area of slidingmode control for nonlinear systems are either based on particular applica-
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
tion areas, such as robotics, or assume that the process satisfies often quiterestrictive structural properties, for example feedback linearisability. It willbe seen that the use of a particular canonical form, the Fliess generalizedcontroller canonical form, enables the existence and reachability problemsto be solved for a relatively broad class of nonlinear systems. It will beshown that the resulting method has the additional advantage of providinga natural way of designing dynamic sliding mode controllers, which effec-tively filter the discontinuous control usually associated with sliding modecontrol methods. The method may also be applied to certain processeswhich are not stabilizable by continuous feedback alone. The use of slid-ing mode control methods involving dynamic feedback has proved to yielduseful results.
5.2 Static output feedback ofuncertain systems
Consider an uncertain dynamical system of the form
x(t) = Ax(t) + Bu(i) + f ( t , x, u)
y(t) = Cx(t) (5.1)
where x e Rn, u e Mm and y e Rp with ra < p < n. Assume that thenominal linear system (A, 5, C) is known, the pair (A, B) is controllableand the input and output matrices B and C are both of full rank. Theunknown function / : R+ x Rn x Em — »• En, which represents the systemnonlinear ities plus any model uncertainties in the system, is assumed tosatisfy the usual matching condition
,u) (5.2)
where the bounded function £ : R+ x Rn x Rm -»• Rm satisfies
Q!(*,y) (5.3)
for some known function a : R+ x W — > R+ and positive constant ki < I .The intention is to develop a control law which induces an ideal sliding
motion on the surface
Q} (5.4)
for some selected matrix F € Rmxp. A control law of the form
u(t) = Gy(t) - vy (5.5)
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
will be sought where G is a fixed gain matrix and the discontinuous vector
1 0 otherwise
where p ( t , y ) is some positive scalar function of the outputs.Consider first the choice of hyperplane to ensure a stable reduced-order
motion. To guarantee the existence of a unique equivalent control
ueq(t) = -(FCB)-lFCAx(t),
it is necessary that det(FCB) ^ 0. It is well known that
rank(FCB) < min{rank(F), rank(OB)}
and so in order for FBC to have full rank both F and CB must have rankm. The matrix F is a design parameter and can be chosen to be of fullrank. A necessary condition therefore for the matrix FBC to be full rankis that rank(CJ3) = m.
The following canonical form will be the key to the developments thatfollow.
Lemma 53 Let (A, B, C} be a linear system with p > m and rank(CB) =m. Then a change of coordinates exists so that the system triple with respectto the new coordinates has the following structure:
a) The system matrix can be written as
A = \ Al1 Al2 Where AU e R(n-m)x(n-m)
and the sub-block AH when partitioned has the structure
0 A22
0 A^22
(5.8)
where A°u 6 Rrxr, A%2 e R(™-P-*-)X(™-P-O andA^ € R(p-™)x(n-P-r)for some r > 0 and the pair (^22 '^21) ^s completely observable.
b) The input distribution matrix has the form
where B2 G ]Rmxm and is nonsingular.
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
c) The output distribution matrix has the form
C=[Q T ] (5.10)
where T £ Rpxp and is orthogonal.
For a proof see [1].Let
p— m m4— >• 4— >
[ F! F2 } =FT
where T is the matrix from equation (5.10). As a result
FC = [ Fid F2 } (5.11)
where
C\ := [ 0(p_TO)x(n_p) I(p-m) ] (5.12)
Therefore FOB = F2B2 and the square matrix F2 is nonsingular. Byassumption the uncertainty is matched and therefore the sliding motion isindependent of the uncertainty. In addition, because the canonical formin Lemma 53 can be viewed as a special case of the regular form normallyused in sliding mode controller design, the reduced-order sliding motion isgoverned by a free motion with system matrix
A'u-^An-AuF^FiCi (5.13)
which must therefore be stable. If K <E R™X(P-™) is denned asK = F2~1Fi
then
Asu = Au-AuKd (5.14)
and the problem of hyperplane design is equivalent to a static output feed-back problem for the system (Au,Ai2,Ci).
In the case where r > 0, the intention is to construct a new system(Aii,Bi,Ci) which is both controllable and observable with the propertythat
X(ASU) = A(Af j ) U A(in - B.KC,)
To this end, partition the matrices A\2 and Afy as
[ 4mJT.1 oi I / f -t f \.i?1 (5.15)A ill I \ /A122
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
Where Am G M(n-m-r)xm an(} ̂ ^ R(n-P-r)x(p-m) ftnd form
sub-system represented by the triple (An, A^,^) where
^11 := /to2 /im2 C*l := [ 0(p_m)x(n-p-r) I(p-m)L -^21 ^22 J
It can be shown that the spectrum of A^ decomposes as
A(An - A12Kd) = A(A^) U A(An - A
and the spectrum of A^ represents the invariant zeros of (A,B,C)[1}. Itfollows directly that for a stable sliding motion, the invariant zeros ofthe system (A, B, C) must lie in the open left-half plane and the triple(An, Ai22, Ci) must be stabilizable with respect to output feedback.
It should be noted that the matrix Ai22 is not necessarily full rank.Suppose rank(Ai22) = m', then it is possible to construct a matrix ofelementary column operations Tm> G ]Rmxm such that
Ai22rm/ = \ BI 0 1 (5.17)L J ^ /
where BI G R(«-™-r)xm' and js of fmj ram< if #m, _ T^}K and Km> ispartitioned compatibly as
Km> =
then
AH — Ai2?KCi = -All — [ BI 0 J Km'Ci = AH — B\K\C\
and (An, Ai22,Ci) is stabilizable by output feedback if and only if thetriple (An, B\,C\) is stabilizable by output feedback. By using PBH testsit can be verified that the pair (An, #1) is completely controllable andthe pair (An,Ci) may be shown to be completely observable [1]. If theKimura-Davison conditions
m'+p + r>n+l (5.18)
are met, the triple (An, jBi,Ci) is stabilizable.Having established conditions to guarantee existence of a stable sliding
motion, a controller to guarantee reachability must now be sought. Assumethere exists a KI G ]Rm/x(p-m) such that AH - BiKiCi is stable. Let
K = T^l"1] (5.19)
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
where K2 e R("»-"»')x(p-m) and ig arbitrary and the matrix Tm> e Rmxm
is defined in equation (5.17). Then providing any invariant zeros are stable,it follows that the matrix A\\ — A\iKC\ is stable. Choose
F = F2 [ K Im]TT
where F^ £ fl£mxm is nonsingular and will be defined later. Introduce anonsingular state transformation x i— » Tx where
-*(n-m) UJl
and C\ is defined in (5.12). In this new coordinate system, the systemtriple (A, B, FC] has the property that
A I -^il -**IZ I 7-1 I U
where A H = A H — A^KCi and is therefore stable. Let P be a symmetricpositive definite matrix partitioned conformably with the matrices in (5.21)so that
9o\0 }
where the symmetric positive definite sub-block P% is a design matrix andthe symmetric positive definite sub-block PI satisfies the Lyapunov equa-tion
PiAn + AfiPi = -Qi (5-23)
for some symmetric positive definite matrix Q\. If
F := B2TP2 (5.24)
then the matrix P satisfies the structural constraint
PB = CTFT (5.25)
For notational convenience let
Q2 := PiAi2 + AjiP2 (5-26)Qs := P2^22 + AJ2P2 (5.27)
and define
7o := iAmax((F-1)T(Q3 + Qjgr1Q2)P-1) (5.28)
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
This scalar is well defined since the matrix on the right is symmetric andtherefore has no complex eigenvalues. It can be shown that the symmetricmatrix £(7) := PAQ + A^P where AQ = A — ̂ BFC is negative definite ifand only if 7 > 7o[l]- A variable structure control law, depending only onoutputs, which will ensure reachability of the sliding mode for appropriatesquare systems is thus given by
u(t] = -
where 7 > 7o and vy is the discontinuous vector given by
if Fy ? 0otherwise
and p(t, y) is the positive scalar function
p ( t , y} = (fci7||Fy \ + a(t, y) + 72) / (1 -
(5.29)
(5.30)
(5.31)
where 72 is a positive design scalar which defines the region in which slidingtakes place. It can be shown [1] that the variable structure control law abovewill quadratically stabilize the uncertain system and a Lyapunov functions
TV(x) := xPx (5.32)
Furthermore an ideal sliding motion is induced on S in finite time.Numerical example
Consider the nominal linear system
00
-1
D C =I
-2
(5.33)taken from [4]. By defining appropriate transformation matrices the systemmay be expressed in the appropriate canonical form as
A =-1.5816 0.0192 0.1457
1.4071 0.3845 -1.70800.2953 0.3400 0.1971
B =00
-3.9016
and/-( _ 0.3417
0.9398-0.93980.3417
It can be verified that B2 = -3.9016, the orthogonal matrix
rji 0.34170.9398
-0.93980.3417
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
and the triple (Au,Bi,Ci) is given by
7 [ -1.5816 0.0192 1 ~ [ 0.1457 1 ^ r ,An = [ 1.4071 0.3845 j BI = [ -1.7080 J & = [ Q 1 \
Here r = 0, hence the original system does not possess any invariant zeros.Arbitrary placement of the poles of AH — B\K\C\ is not possible since onlya single scalar is available as design freedom. For the single-input single-output system (An,B\,C\) the variation in the poles of AH — B\K\C\with respect to K\ can be examined by root locus techniques. In this caseif the gain matrix K = K± = -1.0556, then X(AU - BiKCi) = {-1, -2},from which
F = F2[ K 1 ] TT
= F2 [ -1.3005 -0.6503 ] (5.34)
where F2 is a nonzero scalar that will be computed later. Transforming thesystem into the canonical form using T defined in (5.20) generates
T [ -1.5816 0.172911 ~ [ 1.4071 -1.4184
where \(Au) = { — 1, —2} by construction. It can be verified that
0.3368 0.1891Pl ' 0.1891 0.5401
is a Lyapunov matrix for AH and if P2 = 1, the parameter F2 = —3.9016.It can be checked that 70 = 0.2452 and substituting for F2 in (5.34) gives
F = [ 5.0741 2.5370 ]
The following closed-loop simulation represents the regulation of the initialstates [1 0 0] to the origin. Figure 5.1 represents a plot of the switch-ing function versus time. The hyperplane is not globally attractive since atapproximately 0.3 second it is pierced and a sliding motion cannot be main-tained. Only after approximately 1 sec is sliding established. Figure 5.2shows the decay of the states to the origin.
In summary, for the case of a non-square system, there exists a matrixF defining a surface S which provides a stable sliding motion with a uniqueequivalent control if and only if
• the rank (CB) = m• the invariant zeros of (A, B, C) lie in C_• the triple (An,Bi,Ci) is stabilizable with respect to output feedback.
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
3 4
Time, sec
Figure 5.1: Switching function versus time
t/E 0
-0.5
0 1 2 3 4 5 6 7
Time, sec
Figure 5.2: Evolution of system states with respect to time
The invariant zeros are a property of the system under considerationwhich must usually be regarded as fixed. The next section will explore howa dynamic approach can be used to extend the class of uncertain systemsfor which output feedback sliding mode controllers can be developed. Thiswill be achieved by eliminating the stabilizability restriction.
5.3 Output feedback sliding modecontrol for uncertain systems viadynamic compensation
In the analysis above, it was assumed that the triple (Ai\,B\,C\) wasstabilizable with respect to output feedback. This property can be guaran-teed if the so-called Kimura-Davison conditions hold. If it is not possible
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
to synthesize a K\ to stabilize the triple (An, B\, Ci), then it is natural toexplore the effect of introducing a compensator - i.e., a dynamical systemdriven by the output of the plant - to introduce extra dynamics to provideadditional degrees of freedom.
Consider the uncertain system from equation (5.1) together with a com-pensator given by
xc(t) = Hxc(t) + Dy(t) (5.35)
where the matrices H € Rqxq and D € Rqxp are to be determined. Definea new hyperplane in the augmented state space, formed from the plant andcompensator state spaces, as
Sc = {(x, xc) G Rn+q : Fcxc + FCx = 0} (5.36)
where Fc G Rmxq and F e Rmxp. As in Section 5.2, assume that thenominal linear system (A, B, C) is in the canonical form of Lemma 53 andpartition the matrix FT, where T is the orthogonal matrix from (5.10), as
p—m m<-»• <->•
[ F! F2 ] = FT
In an analogous way define DI e R<?x(p-m) and D2 e M9Xm as
[ Dl D2]=DT (5.37)
If the states of the uncertain system in the coordinates of Lemma 53 arepartitioned as
x = l n~m (5.38)[ X2 J Im V '
then the compensator can be written as
xc(t) = Hxc(t] + DiCiXi(t) + D2x2(t] (5.39)
where C\ is defined in Equation (5.12). Assume that a control action existswhich forces and maintains motion on the hyperplane Sc given in (5.36). Asusual, in order for a unique equivalent control to exist, the square matrix F2
must be invertible. By writing K = F2~1Fi and defining Kc = F^~1FC then
the system matrix governing the reduced-order sliding motion, obtained byeliminating the coordinates x2, can be written as
xi(t) = (An - Ai2KCi)xi(t) - Al2Kcxc(t) (5.40)
xc(t) = (Dl - D2K)ClXl(t) + (H - D2Kc)xc(t) (5.41)
From the above equations it is clear that the introduction of the com-pensator has produced more design freedom. As would be expected, the
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
invariant zeros of the uncertain system are still embedded in the dynamics,since from the definition of the partition of A^ given in (5.15) and froman appropriately partitioned form of AH — A\2KCi, it follows that
-A12KC
H-D2KC
0 Au - A^KCi -Al22Kt
0 (Di - D2K}Ci H-D2K,
As in the uncompensated case, it is necessary for the eigenvalues of A°± tohave negative real parts. The design problem becomes one of selecting acompensator, represented by the matrices Di,D2 and H , and a hyperplanerepresented by the matrices K and Kc so that the matrix
_ I" An - Al22KCi -A122KCAc'- (Di - D2K}d H - D2KC
is stable. Again if there is rank deficiency in the matrix ^122, then theproblem is over-parameterized. As in Section 5.2, suppose rank(Au2) =777,' < ra and let Tm/ £ Rmxm be a matrix of elementary column operationssuch that
Al22Tm, = [ B! 0 ]
where B\ € ]R(n-m-r)XTn and is of full rank. Define partitions of thetransformed hyperplane matrices as
2
then it follows that
An - Bi/^Ci -B,Kclc~~ n- L>2
As before, the matrix given in (5.43) will be written as the result of an out-put feedback problem for a certain system triple. Unfortunately, a degreeof over-parametrization is still present in (5.43), which for simplicity willbe removed by defining
Di := Dl - D2K and H := H - D2KC (5.44)
This is comparable to the situation which occurred in the uncompensatedcase where K2 was found to have no effect on AH — B\KCi. The keyobservation is that Equation (5.43) can now be written as
n-BlKlCl -£iA'ci] = \AU O l l B i 0 1 \K, Kcl]\Cl 0DlCl H \ [ 0 Oj [0 -/,] [Di H
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
Thus by defining
A ._ [ ^n 0 1 R ._ [ Bi 0 1 _ [ Ci 09 '~ n n q'~ n -r \ q '~ n /"L u U«jXg J L U J9 J L U J9
the parameters Ki,Kci,Di, and # can be obtained from output feedbackpole placement of the triple (AQ, Bq,Cq). In order to use standard out-put feedback results it is necessary for the triple (Aq,Bq,Cq) to be bothcontrollable and observable. Prom the definition of (Aq, Bq] it follows that
rank [ zl — Aq Bq ] = rank [ zl — A\\ B\ ] + q
for all z 6 C. As argued earlier, the pair (Au,Bi) is controllable andtherefore from the PBH rank test (Aq,Bq) is controllable. Using the factthat the pair (An, Ci) is observable, a PBH argument proves that (Aq, Cq)is observable.
The Kimura-Davison conditions for the triple (Aq,Bq,Cq] amount torequiring that
m' + q+p>n-r + l (5.45)
Thus for a large enough q, the Kimura-Davison conditions can always besatisfied and the static output feedback method can be employed.
5.3.1 Dynamic compensation (observer based)It is well known that numerical solutions to the static output feedback prob-lem often invoke the use of optimization routines which may not be guaran-teed to converge. This subsection explores an observer-based methodologyfor hyperplane design. Consider the compensator defined in (5.35) then,as in the previous section (eliminating any invariant zeros), the assignabledynamics of the sliding motion are given by the system matrix
A —A " H - D2KC
An alternative method for choosing appropriate compensator variables H , D\and D2, and the hyperplane matrix gains K and Kc will now be sought.
Consider the fictitious system
(t) = Aux(t) + , }
y(t) = Cix(t) { }
with associated triple (An, ^122,^1)- The structure of
C\ = [ 0(p_m)x(n-p-r) I(p-m) \
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
means that the second (p — m)th dimensional component of the 'state' isknown. A reduced order observer would thus only be required to estimatethe first (n — p — r}th dimensional component. If the input distributionmatrix is partitioned conformably so that
A\22l \n-p-r , ,A (0.48;
^1222 J lp-m
then a reduced-order observer for the fictitious system (5.47) is given by
z = (A°22 + L°A°2l)z + (A^22 + L°A%, - (A%2 + L°A02l)L°} y
+ (Ai22i + L°Al222}u (5.49)
where L° e R(n-p-r)x(p-"0 is any gain matrix so that A%2+L°A2l is stable.Let 1C be any state feedback matrix for the controllable pair (An,^122) sothat AH — Ai22/C is stable, and partition the state feedback matrix so that
n — r—p p—m<—> <—>
[ 1C 1C } = 1C
The state feedback law can be implemented using the observer states andthe outputs in the form
u = -ICiz - (JC2 - lCiL°}y (5.50)
and the closed-loop system comprising (5.47) and (5.49) is stable. Define
H — A^2 + L°'A2i (5.51)
D2 = -Al221 ~f" -^ -^-1222 (5.53)
K = /C2-/CiL° (5.54)
Kc = ICi (5.55)
then equation (5.49) can be written
z ( t ) - Hz(t) + Diy + D2u (5.56)
whereu = -Kcz(t) - Ky (5.57)
It can easily be verified that the closed-loop system formed from (5.47) and(5.49) is given by
AH - A^KCi -Ai22Kc 1 \ x(t]z ( t ) \-[(D,- D*K)Ci H - D2KC ' (5'58)
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
and from the separation principle the closed-loop poles are given by
X(H)\jX(An-A122lC)
The system matrix associated with (5.58) is identical to the system matrixof the reduced-order sliding motion given in (5.42). Therefore the choiceof compensator matrices in (5.51) to (5.53) and the hyperplane matrices(5.54) and (5.55) give rise to a stable sliding mode.
5.3.2 Control law construction
Having investigated design procedures to determine the compensator andassociated sliding surface, it is necessary to construct a control which willrender the defined sliding mode attractive. Assume that there are r (stable)invariant zeros and partition the state vector x\ as in (5.38) so that
xu ln-p-r
Ip-m
As a result, the (original) compensator can be written as
xc(t] = Hxc(t) + Dixi2(t) + D2x2(t) (5.59)
Define a new dynamical system by
zr(t) = A°uzr(t) + A°l2xc(t) + (A?2l - A°l2L°)xl2(t) + Al21x2(t) (5.60)
and augment (5.59) with (5.60) to form a new compensator
xc(t) = Hxc(t) + Dy(t) (5.61)
where
A°12 1 ._ (Af2l - A°12L°) Al2l
0 H - D, D2T
T
Using the partitions (5.7), (5.8), (5.15) and (5.48), the original dynamicscan be written as
xr(t) = A^xr(t) + A°2xn(t) + A?2lx12(t} + A121x2(t) (5.62)
n(0 = A°22xu(t) + A™22x12(t) + Al221x2(t) (5.63)
i2(t) = A^xu(t) + A£x12(t) + Al222x2(t) (5.64)
X2(t) = A2nXr(t) + A2i2Xu(t) + A2i3Xi2(t)+A22X2(t)
(5.65)
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
where the lower left sub-block of A from (5.7) has been partitioned so that
r n—p—r p—m
r 7 ," „" i A <5'66)I ^211 ^212 ^213 j = ^21
Define two error states
and
Arr JUr
ec — xc — xn — L
then straightforward algebra reveals
e r ( t ) = A°uer(t) +
and alsoec(t) = Hec(t)
(5.67)
(5.68)
(5.69)
(5.70)
These stable error systems result from the fact that, by construction, thecompensator states xc and zr are observations of xu + L°x\2 and xr, re-spectively. Define a state matrix
x =
X2
then standard algebra reveals
x(t] = Ax(t) - Aee(t) + B \u(t) + £(t, x, u)}
where
(5.71)
(5.72)
A°u A°12 A?2l-A°l2L°O H DI0 A§! A%-A°21L°
^211 -^212 ^213 ~ -^212-^°
0 00 00 A°2
A211 A2i
D and
and the augmented error state
e = (5.73)
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
Note that the triple (A, B, C) can be obtained from the canonical form(A, B, (7) via a similarity transformation. Thus the original system togetherwith the compensator can be written as
k(t) = He(t) (5.74)
x(t] = Ax(t}-Aee(t) + B[u(t)+£(t,x,u)} (5.75)
Note also that the sliding surface Sc can be written as
[x <E Rn : Sx = 0}
whereS = F2 [ Omxr Kc K Im } (5.76)
Define a switching function
s(t) = Sx(t) (5.77)
and define a linear feedback component
ui(t) = -A-lSAx(t) + A~l3>Sx(t) (5.78)
where A = SB and $ e £mxm jg a stable design matrix. Let P be theunique positive definite solution to the Lyapunov equation
P$ + $TP = -/ (5.79)
A control law to induce a sliding motion on the sliding surface <5C is givenby
u(t) = ui(t) - vy (5.80)
where
if .(0*01 0 otherwise
and p(-) is the positive scalar function
p(t, y) = (*i || A|| \\ui(t) || + || A||a(«, y) + 72) / [1 - fci«(A)] (5.82)
where 72 is a small positive constant.By considering a Lyapunov candidate of the form V(s) = sTPs where
s(t) = 5x(t), it may be shown that the control law defined in (5.78) to(5.82) induces a sliding motion on the sliding surface Sc.
This control law is effectively a state feedback controller since the com-ponents zr and xc are estimates of the true states xr and x\\ (up to acoordinate transformation) .
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
5.3.3 Design exampleConsider the nominal linear system
A =
-2001
I 0 00 4 11 0 0
-6 -9 -2
D c = 0 0 0 10 0 1 0
This system is already in the appropriate canonical form and thus
ln L121J-2
00
1 00 41 0
(5.83)
0
0 Au
In terms of the compensator design, the triple of interest is given by
^122 = [ J ] Ci = [ 0 1 ] (5.84)
It can be shown by direct computation that for K = k
J o
and so the triple (An, ^122, Ci) is not stabilizable by static output feedbackand a compensator-based approach must be employed. It follows from(5.84) that
^22 122 _
22
and so from Equations (5.51) to (5.53) an appropriate parametrization forthe compensator is
H = L° ° 2= 4 - (L°) = 1
where L° is any negative scalar which will appear as one of the eigenvaluesof (5.42). In the simulation which follows L° — —2.5 and A(^4n —^122^) ={ — 1, —1.5}. Since the system has an invariant zero at —2, the sliding mo-tion will have poles at { — 1, — 1.5, — 2}. The pole represented by $ whichgoverns the range space dynamics has been chosen to be —5. For simplicitythe scaling factor for the sliding surface is ^2 = 1. All the available degreesof freedom have now been assigned.
Figure 5.3 is a plot of the switching function against time; it can beseen that sliding occurs after approximately 1 second. Figure 5.4 shows the
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
.1 0-4
UH60
2 °-2
5
Time
10
Figure 5.3: Switching function versus time
5
Time
10
Figure 5.4: Evolution of the system states
evolution of the states against time. Initially the states of the compensatorhave been set to zero. The states of the system have a nonzero initialcondition which needs to be regulated to zero.
Figures 5.5 and 5.6 show the evolution of the error states ec and er. Ini-tially ec is nonzero since the state x\\ was given a nonzero initial condition.As indicated in Equation (5.70), this error system is completely decoupledand decays away to zero (Figure 5.5).The error states er, shown in Figure 5.6, although initially zero, are cou-pled to the state ec as shown in Equation (5.69). However, this also decaysasymptotically to zero in accordance with the theory.Notice from Figure 5.3 that, although the states initially lie on the slid-ing surface, a sliding motion is not maintained. This is due to the factthat the error term e is initially too large. A sliding motion occurs afterapproximately 1 second, by which time the error e has decayed sufficiently.
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
8 -0.4
-0.8
5
Time
10
Figure 5.5: Evolution of the error states ec
5Time
10
Figure 5.6: Evolution of the errors states er
5.4 Dynamic sliding mode control fornonlinear systems
Sliding mode control is known to provide an appropriate solution to therobust control problem. However, the majority of design methodologies,whether reliant on state or output feedback, have been based around linearuncertain systems, as described earlier in this chapter, or specific typesof nonlinear systems. The latter may involve particular application areas,such as robotics [10], or require that relatively stringent conditions are metby members of the system class: for example the system class may be re-quired to be feedback linearizable [11]. It is obviously desirable to have asliding mode control methodology that will be applicable to a fairly broadclass of nonlinear system representations, exhibit robustness while yieldingappropriate performance, and lend itself to the development of appropriate
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
tool boxes for controller design. It will be shown in the remainder of thechapter that the dynamic sliding mode policies which result from consid-ering differential input-output (I-O) system representations are sufficientlygeneral to meet this remit [5, 6, 9].
Dynamic sliding mode control methods assume that all the systemsstates, or equivalently, the derivatives of the outputs to some appropriateorder, are available for use by the control law. Thus a state estimator isnecessary for implementation if only measured outputs are available.
The following notation will be used throughout:
NS(XQ) = {x e Rn : \\x-x0\\<6}
where 6 > 0, or simply N& if XQ — 0.For sliding mode controller design using static feedback, it is necessary
that the system assumes a regular form and that the control variables ap-pear linearly in the system in order to recover the control parameters fromthe chosen sliding condition [13]. In general, this is not practically imple-mentable for general nonlinear systems with nonlinear control. In order todevelop the sliding mode control method to include dynamic policies, andhence to ensure it becomes applicable to an extended class of nonlinearsystems, differential I-O system representations will be employed.
For a given system in state-space form that is locally observable,
x = /(x,M) , .y = Mx,u,*) (5'85)
where x € Rn ,u € Rm, and /(x, u) and /i(x, u) are smooth vector func-tions, the following locally equivalent differential I-O system exists [14]
(5-86)(Tin)
ypwhere
u = (Ul, . . . , u(0l\ . . . ,um, . . . ,^m))T (5.87)
andy = (yi,. . . ,i/in i~1) , . . . ,yP , . . . ,^n ' -1))T (5.88)
with HI + . . . + rip = n.
Definition 54 A differential I-O system (5.86) is called proper if
1) p = m;
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
2) All ( /?;( . , . , . ) , i = 1 , . . . , ra, are C3 functions;
3) Regularity condition
det (5.89)
is satisfied with y e N$(Q) for all t > 0, some 5 > 0 and genericallyfor u.
Throughout this chapter it is assumed that all the differential I-O sys-tems considered are proper.
Whether or not the resulting system is minimum phase will again beshown to be pertinent to the stability of the closed-loop system.
Definition 55 The zero dynamics, corresponding to (5.86), is defined as
¥>i(0,M = 0: (5.90)
y?p(0,iM) = 0.
The system (5. 86) is called minimum phase if there exist 5 > 0 and UQ € M'3
where 0 = /3\ + . . . + /3m, such that (5.90) is uniformly asymptotically(exponentially) stable for an initial condition u(0) £ A^(UQ), where
Otherwise, it is non-minimum phase. Note that, in this case, the "mini-mum phase-ness" is a property of the chosen control signal.
In order to address robustness, uncertain systems of the following formmay be considered.
: (5.91)
y(pp} - vP(y,u,t) + Ap(y,t)The uncertainties are Lebesgue measurable and satisfy
| |A i (y , t ) | |<p i | |y | |+/ i , P i > 0 , / < > 0 , i = l , . . . , p (5.92)
The uncertainty may be due to external uncertainties, internal param-eter uncertainties, measurement noise, system identification error, or in-deed the elimination procedure used to generate a differential input-outputmodel from a state space model as in [14].
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
It is often convenient to consider the Generalized Controller CanonicalForm (GCCF) representation of (5.86). Without loss of generality, supposethat n i , . . . ,nmi > 1 and nmi = . . . = nmi+m2 = l,mi + ra2 = m. Thesystem (5.86) may be expressed in the following GCCF [2]
Xi) _ Xi)M — S.2
XI) _ XI)Sni-1 ~ VH
Xm) _ X™)Snm-l — Snm
C'(TO) / i. ^ ,\L = <pm(C,U,*)
whereXi) _ (X») Ai)\ _ ( i,(ni~l)} i-l mS> — VSl » • • • ) Sni J ~~ Vi/i) • • • ) Ui ;, t — 1, . . . , »t
and
(5-93)
represent the system outputs and their derivatives.It has been seen that it is necessary to solve existence and reachability
problems in order to determine the sliding mode controller. In the nonlinearcase, the two popular choices of sliding surface are:
(1) Direct sliding surface [8, 6, 9]
, » = , . . . , m3 = 1
where Y^j=i aj ; •^J'~1 are Hurwitz polynomials with anl = 1. This willprovide a reduced order sliding motion whose dynamics are prescribedby the roots of the polynomials.
(2) Indirect sliding surface [5]
+<^(C,<M), t = l , . . . ,m (5.95)
where X^j^t aj ^~1 are Hurwitz polynomials with a^+1 = 1. Withthis choice, the system (when sliding) becomes equivalent to an nthorder linear system, with dynamics prescribed by the choice of theHurwitz polynomial. This may be regarded as an alternative model.
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
An appropriate algorithm for robust dynamic sliding mode control isdescribed below. The system (5.91) may be expressed in the followinggeneralized controller canonical form
Ci = (2
(5.96)
where
HO _ (XO H0\ _ (7/. ?/(ni~1h i - 1
S — VSl 5 • • • ) Sn, ; ~~ V^' ' ' ' ' i/i ;, t — 1, . . . ,
and
Step 1: Choose design parameters to define the sliding surface (5.94). For
i = 1, . . . , m, if m > 1, choose (a^\ . . . , aj^.j, 1) and (a^, . . . , ai,_i),both Hurwitz. This is always possible according to the result in [3]. With-out loss of generality, suppose HI, . . . , nmi > 1 and nmi+i = . . . = nmi+m2
where mi + m-2 — m-Step 2: Estimate the uncertainty bound as in (5.92) when the system isin the GCCF. Choose 00 and 0 where 0 < 0 < 1, 00 + 0 = 1 and define
i=l
where
(5.98)
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
Step 3: Define
0 1 0 0
0 0 1 0
0 0 0 1
0
for i = 1, . . . , mi. Let D := diag[D\, . . . ,Dmi] and A := diag[A\, . . . , -AmJ.Since the A{ are stable, then A is stable, and define P to be the uniquepositive definite solution to the Lyapunov equation
ATP + PA = -I
Next choose K e ]RmXTn as a positive definite matrix which satisfies
plmi > 0 (5.99)
Step 4: Differentiating (5.94) with respect to time t along the trajectoriesof (5.96) leads to
(5.100)
(5.101)
5'i 1(5-96) = ] o Cj=l
for i = 1, . . . , m. Now set
J=l
where fc0i > IQ := lA2)1/2and
x { *• X>£i t \ i / \ J 1 1 ^ -
QS1T I T* 1 —— £ " • C/TFTf 1 / T* T* ^ JTO(JjLg\JU } C OCl»6l ) \ X) |»// \ C
£ ( -1, X < -£
Equation (5.100) becomes
s = -Ks - K0sat£(s) + A(C, t) (5.102)
where KQ = diag(kQ\,..., fcom] , sate(s] = [sate(si),..., sate(sm)]T.
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
Step 5: From (5.101) the highest order derivatives of the control,namely [u\ , . . . ,Um ], can be solved out by the implicit function the-orem as
u(0i] =p i (C ,u , t ) , i = l , . . . , m
if the regularity condition is satisfied. Note that pi(£,u, t) is a continuousfunction if Si ^ 0 because (f>i is C1 and 7$ is C° if Si ^ 0. This dynamicfeedback can be realized in canonical form by introducing the pseudo-statevariables as
(5.103)
• (TO) (m)Z \ / - *y V /
— —
where
~(i) _ ( J*) r^h — (•? / . V i - i/^"1^ ? — 1 mz — \z^ , . . . , Zp. ) — ( u l ^ u l ^ . . . ̂ ui j , i — i, . . . , m ^
and2-(^ ( 1 ) , . . . ,2 ( m ) ) T . (5.105)
The system in (5.103) together with (5.96) yields a closed-loop system ofdimension X^Li n» + Y^iiLi &i, where fa is the highest order derivative ofWj.
Step 6: Choose UQ E ffi^ and a 5 > 0 such that, for initial conditionu(0) 6 A^(u0), and
(1) the regularity condition is satisfied;(2) the zero dynamics (5.90), (or (5.103) when C — 0) are uniformly
asymptotically stable; and(3) all the initial conditions for (5.96, 5.103) are compatible.It was shown in [6] that the procedure outlined above will effect uni-
formly ultimately bounded motion of the uncertain system (5.91) if it isminimum phase.
RemarksThe proof in [6] relies on first showing that the closed-loop subsystem asso-ciated with the states (, is stable. This is demonstrated by considering the
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
system (£, s) obtained from the linear coordinate transformation resultingfrom substituting for Qv according to the formula
[which is a rearrangement of (5.94)]. Defining (^ := (Ci , • • • >Cni-i)T ^follows that
C(i) = AiC + Asi (5.106)
for i = 1, . . . , mi. Using the candidate Lyapunov function
ultimate boundedness of the (£, s) subsystem, with respect to an arbitraryneighborhood of the origin, can be shown. The overall closed-loop systemis given by (5.106) and (5.102) together with equations of the form
* = »KC,*,M) (5-107)
where the right hand side is such that z — 77(0, 0, 2, t) represents the zerodynamics. Using the stability properties of the states s and £, stability ofthe overall closed-loop system can be shown by using a modification of theresults for 'triangular systems' in [15].
Equation (5.102) can be shown to represent a strong reachability condi-tion in the sense of [12].
The dynamic sliding mode control method above assumes that all thesystem states, or equivalently, the derivatives of the outputs to some appro-priate order, are available for use by the control law. Thus a state estimatoris necessary for implementation if only measured outputs are available. Aparticular high gain observer was shown to be particularly appropriate forthis estimation task [7].
5.4.1 Design example
Consider the following nonlinear model
A = ysm(y} + rand(l) (5.109)
Here A represents the uncertainty and rand(\] is the one dimensional ran-dom variable from MATLAB. The corresponding zero dynamics are obtainedby setting y^ — y^ = y = 0. This yields
w ^ + w + yuu ( 1 ) (w 2 - l ) = 0 (5.110)
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
which is the Van der Pol equation. This is uniformly asymptotically stablefor LI < 0 with [w(0)j + [^^(0)] < 1 as shown by the phase plane portraitin Figure 5.7 where JJL — — I and u(0) = u(0) = 0.5. The system is thusminimum phase and the closed-loop dynamic sliding mode control schemewill be stable for appropriately chosen initial parameters.
0.2 0.3 0.4 0.5-0.4
-0.2 -0.1 0.6 0.7
Figure 5.7: Phase plane portrait showing the typical evolution of the zerodynamics
Step 1: Choose the direct sliding surface s = ay -f y^ where a = 2.Step 2: To estimate the uncertainty bound, choose OQ = 0.25 so that
6 = 0.75. From |A| < |y| + 1, it follows that /0 = 1 and p^ = 3 impliesp = 4.
Step 3: As A = -2, it follows that P = 0.25. Thus
(5.111)
and appropriate choice for k in the reachability condition is k = 5 > 4.25.Let &o = 1.5 > /o = 1.
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
Step 4: The controller is then solved out from s = —ks — kosat£(s) as
UW = -(ks+k0sat£(s)+usm(y(l)}+uy+u+nu(l)(u2-l)}/(l+y2) (5.112)
Simulation results for initial conditions y(0] = 0, y(0) = 0.5 are illus-trated. The level of uncertainty present is shown in Figure 5.8. The outputresponse is shown in Figure 5.9. It is seen that the closed-loop systemrejects the uncertainty and effective output regulation is achieved. Fig-ure 5.10 shows that a sliding mode is attained and maintained. It is seenfrom Figure 5.11 that this performance is achieved with out the switchedcontrol action, which is often associated with sliding mode control. Thedynamic control strategy act as a natural filter for the control signal andits robustness to the prescribed uncertainty results.
o 0.5
Figure 5.8: Evolution of the uncertainty contribution to the dynamics
o 0.5 i
Figure 5.9: Evolution of the system output with respect to time
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
S n i so§ n 1oo.5 n CKo
'% n00 U
.fins0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time
Figure 5.10: Evolution of the switching function
0.04
-0.060 0.5 1
Figure 5.11: Evolution of the control signal with respect to time
5.5 Conclusions
In this chapter design procedures have been presented to synthesize ro-bust output feedback controllers for linear uncertain systems. The class ofsystems to which the results apply has been identified, and includes therequirement that the nominal linear system is minimum phase. It has beenshown that certain dimensionality requirements must be satisfied if the slid-ing surface is to be designed using a straightforward static output feedbackpole placement, which is dependent only on a particular subsystem of theoriginal plant dynamics. This restriction can be overcome using a dynamicfeedback approach. A reduced-order Luenberger observer approach wasshown to yield a convenient methodology for designing the sliding surfaceand compensator dynamics in this case. An output dependent controllerwhich guarantees attainment of a sliding mode by the linear uncertain sys-tem was presented.
This chapter also addressed the problem of designing sliding mode con-
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
trailers for nonlinear systems. A particular canonical form was used to ren-der the results applicable to a fairly broad class of systems. This methodwas shown to produce controllers that are dynamic in nature and thusavoid the chattering which often characterizes the sliding mode approach.
References[1] C. Edwards and S.K. Spurgeon, "Sliding mode control: theory and
applications", Taylor & Francis, London, 1998.
[2] M. Fliess, "What the Kalman state variable representation is goodfor", Proc. IEEE CDC, Honolulu, Hawaii, pp. 1282-1287, 1990.
[3] W. Hahn, "Stability of Motion", Springer-Verlag, New York, 1967.
[4] S. Hui and S.H. Zak, "Robust output feedback stabilisation of uncer-tain dynamic systems with bounded controllers", International Jour-nal of Robust & Nonlinear Control, 3 , pp. 115-132, 1993.
[5] X. Y. Lu and S. K. Spurgeon, "Asymptotic feedback linearisation ofmultiple input systems via sliding modes", Proc. 13th IF AC WorldCongress, San Francisco, pp. 211-216, 1996.
[6] X. Y. Lu and S. K. Spurgeon, "Robust sliding mode control of uncer-tain nonlinear systems", Systems & Control Letters, 32 , pp. 75-90,1997.
[7] X. Y. Lu and S. K. Spurgeon, "Output stabilisation of MIMO non-linear systems via dynamic sliding mode", International Journal ofRobust & Nonlinear Control, 9, pp. 275-306, 1999.
[8] X. Y. Lu, S. K. Spurgeon, and I. Postlethwaite, "Robust variable struc-ture control of PVTOL aircraft", International Journal of SystemsScience, 28, pp. 547-558, 1997.
[9] H. Sira-Ramirez, "On the dynamical sliding mode control of nonlinearsystems", International Journal of Control, 57, pp. 1039-1061, 1993.
[10] J. J. Slotine and S. S. Sastry, "Tracking control of nonlinear systemsusing sliding surfaces, with application to robot manipulators", Inter-national Journal of Control, 38, pp. 465-492, 1983.
[11] S. K. Spurgeon and R. Davies, "Nonlinear control via sliding modesfor uncertain nonlinear systems", CTAT Special Issue on Sliding ModeControl, 10, pp. 713-736, 1994.
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.
[12] S.K. Spurgeon and X.Y. Lu, "Dynamic sliding mode control designusing symbolic algebra tools", In N. Munro, editor, Symbolic methodsin control system analysis and design, Peter Peregrinus, Stevenage UK,pp. 227-250, 1999.
[13] V. I. Utkin, "Sliding Modes in Control and Optimization", Springer-Verlag, Berlin, 1992.
[14] A. J. van der Schaft, "Representing a nonlinear state space system asa set of higher-order differential equations in the inputs and outputs",Systems & Control Letters, 12, pp. 151-160, 1989.
[15] M. Vidyasagar, "Decomposition techniques for large-scale systemswith nonadditive interactions: stability and stabilizability", IEEETransactions on Automatic Control, AC-25, pp. 773-779, 1980.
Copyright 2002 by Marcel Dekker, Inc. All Rights Reserved.